Functional and Parametric Estimation in a Semi- and Nonparametric Model with Application to Mass-Spectrometry Data

Weiping Ma, Yang Feng, Kani Chen, Zhiliang Ying

TL;DR

This paper introduces a novel semi- and nonparametric model to analyze mass spectrometry data, effectively capturing individual variations in protein abundance while identifying a common underlying spectral shape. Their proposed multi-step iterative algorithm simultaneously estimates both the individual-specific parameters and the shared spectral shape, achieving optimal convergence rates and demonstrating strong performance on real and simulated data.

What Problem Does This Paper Address?

Mass spectrometry (MS) generates complex data that reveals proteomic patterns related to biological samples. However, analyzing this data is challenging due to several factors:

  • Individual Variability: Each biological sample (e.g., a patient) has unique characteristics, leading to variations in protein abundance (heights of spectral peaks) and potentially shifts in their measured mass-to-charge ratio (m/z).
  • Common Underlying Structure: Despite individual differences, there's an underlying common spectral "shape" or pattern that is shared across samples, representing the fundamental proteomic features.
  • Data Preprocessing Challenges: MS data often requires extensive preprocessing, including noise filtering, baseline correction, and alignment, to make it amenable to analysis.
  • Existing Model Limitations: Traditional statistical models struggle to simultaneously account for both the individual-specific location/scale parameters and the shared, unknown nonparametric functional form of the spectral shape.

This paper aims to develop a statistical framework that can effectively model and analyze this type of data by disentangling the individual-specific components from the common, non-parametrically modeled spectral shape.

Methods

The core of the paper is a multi-step iterative estimation procedure designed to tackle the semi- and nonparametric model:

  • The Model: The proposed model is $Y_{it} = \alpha_i + \beta_i m(x_{it}) + \sigma_i(x_{it})\epsilon_{it}$, where:
  • $Y_{it}$ is the observed response (intensity) for individual $i$ at time $t$.
  • $x_{it}$ is the mass-to-charge ratio (m/z) value.
  • $\alpha_i$ and $\beta_i$ are individual-specific location and scale parameters, respectively.
  • $m(\cdot)$ is a common, unknown nonparametric regression function representing the spectral shape.
  • $\epsilon_{it}$ are error terms.
  • Identifiability: To address identifiability issues (where $\alpha_i, \beta_i, m(\cdot)$ can be confounded), the authors fix $\alpha_1 = 0$ and $\beta_1 = 1$ for the first individual, effectively using their spectrum as a baseline.
  • Iterative Estimation: The algorithm alternates between estimating the parametric ($\alpha_i, \beta_i$) and nonparametric ($m(\cdot)$) components:
  • Initial $m(\cdot)$ Estimation: Using the first individual's data (with $\alpha_1=0, \beta_1=1$), an initial estimate of $m(\cdot)$ is obtained using local linear regression.
  • Parametric Estimation: With the estimated $m(\cdot)$, the parameters $\alpha_i$ and $\beta_i$ for $i \ge 2$ are estimated using least squares.
  • Updated $m(\cdot)$ Estimation: Using the estimated $\alpha_i$ and $\beta_i$, the estimate of $m(\cdot)$ is updated again using local linear regression.
  • Iteration: Steps 2 and 3 are repeated until the estimators converge.
  • Local Linear Regression: This is a standard nonparametric technique used for estimating the smooth function $m(\cdot)$. It involves fitting local linear models to estimate the function's value and slope at each point.
  • Bandwidth Selection: The performance of local linear regression heavily relies on the choice of bandwidth ($h$). The paper discusses optimal bandwidth selection for both the initial and updated $m(\cdot)$ estimates, aiming to achieve optimal convergence rates. Cross-validation is proposed as a practical method for data-driven bandwidth selection.

Datasets

  • Real Data: The method is applied to a SELDI-TOF mass spectrometry dataset from a study of 33 liver cancer patients. This dataset was collected at Changzheng Hospital in Shanghai.
  • Simulated Data: The authors also conduct simulation studies to evaluate the performance of their method. The true curve $m(\cdot)$ for simulations is derived from a moving average smoother of a fraction of the real mass spectrometry data.

Key Insights and Ideas

  • Decoupling Individual and Common Effects: The core innovation is the ability to disentangle individual-specific location and scale parameters from a shared, unknown spectral shape. This is crucial for understanding both individual differences and common biological patterns in MS data.
  • Multi-Step Iterative Approach: The proposed iterative algorithm is a clever way to handle the interdependence between the parametric and nonparametric components. By alternating estimation, it allows for consistent estimation of both parts.
  • Optimal Convergence Rates: The theoretical analysis shows that the proposed method can achieve optimal convergence rates for both the parametric and nonparametric estimators, provided the bandwidths are chosen appropriately. This is a significant theoretical contribution.
  • Bandwidth Sensitivity: The paper highlights that the choice of bandwidths for the nonparametric components is critical and that different bandwidths might be optimal for different stages of the estimation process.
  • Practical Applicability: The method is demonstrated on real MS data, showing its ability to capture key features like peak shapes and relative heights, suggesting its utility in biological and clinical applications.

Key Results

  • Consistency and Asymptotic Normality: The authors prove that their estimators for both the parametric ($\alpha_i, \beta_i$) and nonparametric ($m(\cdot)$) components are consistent. The parametric estimators are also shown to be asymptotically normal.
  • Optimal Nonparametric Rate: The nonparametric estimator $m(\cdot)$ achieves the optimal rate of convergence when the bandwidths are appropriately selected.
  • Real Data Application: The method was applied to a liver cancer patient dataset, and the estimated baseline function captured the majority of the peaks and shapes observed in the raw data.
  • Simulation Performance: Simulation studies showed that the multi-step procedure provides accurate estimation of parameters and the spectral curve, with performance comparable to an oracle estimator (where the true function is known).
  • Bandwidth Selection: Cross-validation was used to select optimal bandwidths, and the results indicated that the optimal bandwidths for the initial and updated $m(\cdot)$ estimates were different, aligning with theoretical expectations. For the real data, optimal bandwidths were found to be $h=34$ and $h^*=4$.

Limitations and Future Work

  • Assumptions: The theoretical results rely on certain regularity conditions, such as the continuity and bounded second-order derivative of the baseline intensity function $m(\cdot)$, and the density of $x_{it}$ being bounded away from zero. While reasonable for MS data, these might not hold in all scenarios.
  • Measurement Error in X: The proposed model does not account for potential measurement error in the $x_{it}$ (m/z) values. The authors suggest incorporating measurement error models as a future direction.
  • Computational Efficiency: While the iterative process often converges quickly, for very large datasets, computational efficiency might still be a consideration.
  • Extension to Classification: The authors note that the extracted peak information from their method could be used for disease classification, suggesting this as a potential future application.
  • Alternative Estimation Techniques: The paper mentions that local linear fitting could be replaced by other nonparametric techniques, and empirical Bayes or nonparametric Bayes methods could be explored for estimating subject-specific parameters.

Summary updated

Abstract

Motivated by modeling and analysis of mass-spectrometry data, a semi- and nonparametric model is proposed that consists of linear parametric components for individual location and scale and a nonparametric regression function for the common shape. A multi-step approach is developed that simultaneously estimates the parametric components and the nonparametric function. Under certain regularity conditions, it is shown that the resulting estimators is consistent and asymptotic normal for the parametric part and achieve the optimal rate of convergence for the nonparametric part when the bandwidth is suitably chosen. Simulation results are presented to demonstrate the effectiveness and finite-sample performance of the method. The method is also applied to a SELDI-TOF mass spectrometry data set from a study of liver cancer patients.
BibTeX
@article{ma2015functional,
  author = {Weiping Ma and Yang Feng and Kani Chen and Zhiliang Ying},
  title = {{Functional and Parametric Estimation in a Semi- and Nonparametric Model with Application to Mass-Spectrometry Data}},
  journal = {The International Journal of Biostatistics},
  year = {2015},
  doi = {10.1515/ijb-2014-0066},
}
Figure